A single .kicad_sch was parsed in isolation, so child sheets looked like an empty netlist. Report 404s now surface the real API error instead of a generic fetch failure.
Co-authored-by: Cursor <cursoragent@cursor.com>
Pipeline samples PCB traces when stackup εr/h is present; power/ground are skipped. Named nets can be re-analyzed from the Impedance tab.
Co-authored-by: Cursor <cursoragent@cursor.com>
PS-PLC-004 is a foreign net endpoint inside the KiCad courtyard. The lista↔codice table and sprints 0–10+ get an OK column where the hole is closed without invented millimetres.
Co-authored-by: Cursor <cursoragent@cursor.com>
X1 uses the same proximity rule as ICs. A detour is shortest path on PCB segments versus that millimetre, not a guessed loop ratio.
Co-authored-by: Cursor <cursoragent@cursor.com>
PS-PLC-003 uses the same_layer flag plus F/B layers from the PCB. A via in the courtyard is enough; an unset flag skips.
Co-authored-by: Cursor <cursoragent@cursor.com>
Skip without courtyard geometry or a numeric count — no pad-radius default. Tests use simple_project plus explicit coordinates as parameters.
Co-authored-by: Cursor <cursoragent@cursor.com>
PS-SI-001 and decoupling skip without a numeric layout_rule. Tests use simple_project nets, not invented USBPHY boards.
Co-authored-by: Cursor <cursoragent@cursor.com>
PS-PLC-001 fires only with a LayoutGraph: 15 mm vs a 2 mm rule is an error, a 1 mm cap is silent, and a null millimetre uses the declared 3 mm default as WARNING.
Co-authored-by: Cursor <cursoragent@cursor.com>
Accepted findings become ECO rows with a required reason; false-positive and wontfix stay off the change list. Pipeline re-runs keep dispositions by finding_id like comments.
Co-authored-by: Cursor <cursoragent@cursor.com>
Use the vendored Hammerstad-Jensen/Cohn engine for microstrip, stripline, and coupled-diff advice. Skip OpenEMS/pcbnew and refuse CPWG rather than inventing a number.
Co-authored-by: Cursor <cursoragent@cursor.com>
PROVIDER_*=anthropic and Claude fallbacks now coerce to DeepSeek, skills stay local, and default_model_version bumps without upload_skills.py.
Co-authored-by: Cursor <cursoragent@cursor.com>
Cache distributor EOL/NRND/RoHS without inventing replacements, flag catalogued errata pull-ups only, and store layout_rules with a closed kind enum so millimetres stay null unless the datasheet stated a number.
Co-authored-by: Cursor <cursoragent@cursor.com>
Write pinscope-findings.json from the report, keep symbol uuid and child-sheet on the graph, and let the pcbnew plugin jump to a footprint or show the schematic uuid when IPC focus is unavailable.
Co-authored-by: Cursor <cursoragent@cursor.com>
Compare only specified IQ plus I_load to Iout_max and series-R drop, flag sequencing only when power_sequence is in specs, and treat DNP as a fitted-variant graph so a missing enable pull is ERROR only when the BOM actually marks DNP.
Co-authored-by: Cursor <cursoragent@cursor.com>
Match RC/LC/π/T without inventing fc, compare to ADC rate only when specs list it, and skip ferrite DCR unless the IC gives a limit. LDO Tj uses I_load not Iout_max, and missing θJA stays INFO.
Co-authored-by: Cursor <cursoragent@cursor.com>
Keep these as WARNING hints with wide bounds so a 4.7 kΩ or 100 nF VDD cap is not treated as a datasheet µF error.
Co-authored-by: Cursor <cursoragent@cursor.com>
V4.1 is natively multimodal so every stage uses deepseek-flash; pricing and the UI now surface dollars instead of empty credits. KiCad netlists and neighborhood fingerprints land in the same cut so a second run can keep the project and skip unchanged ICs.
Co-authored-by: Cursor <cursoragent@cursor.com>
Ignore enable straps on a power rail, NC nets, and SPI aliases; treat FB as an inductor and RN as a resistor so pull-ups and beads classify correctly.
Co-authored-by: Cursor <cursoragent@cursor.com>
Skip the model (and the credit gate) when LCSC/DigiKey already list C/V/Z, when the BOM value is a single R/C/L/FB token, or when the MPN is an LQW18AN inductor. Ferrite beads keep impedance, not invented henries; value-only specs stay out of the shared library.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add Microchip/Murata/Silicon Labs URL tables, optional Mouser search, and suggested URLs on a miss so the wizard can open PDFs in the user's browser when the VPS is blocked.
Co-authored-by: Cursor <cursoragent@cursor.com>
Hub Audio pull-ups sit on 3V3_DIGITAL typed as signal, and ADAU/TAC5212 SPI pins list SDA in the name — both produced false I2C/reset warnings.
Co-authored-by: Cursor <cursoragent@cursor.com>
Flag supply nets with no cap to ground and open-drain I2C/reset nets with no pull-up, without guessing datasheet capacitor values.
Co-authored-by: Cursor <cursoragent@cursor.com>
DeepSeek review/extraction was using raw get_text; sparse pin drawings now get vision slots and quote checks share the same page text.
Co-authored-by: Cursor <cursoragent@cursor.com>
Follow HTML interstitials to a matching .pdf, send a real Referer, retry http as https, and probe TI gpn plus Espressif/ST/ADI/NXP/onsemi paths when LCSC is blocked.
Co-authored-by: Cursor <cursoragent@cursor.com>
ERROR/WARNING without a real citation are demoted to Unverified so invented topologies like a unidirectional TPD2E007 clamp cannot ship as ERROR.
Co-authored-by: Cursor <cursoragent@cursor.com>
Those 13 U* were in the BOM (TPS22965, TPD2E007, TMP117, …) but never entered ic_mpns, so review skipped them as missing PDFs. Also try the BOM Datasheet URL first.
Co-authored-by: Cursor <cursoragent@cursor.com>
A docker rebuild left HubAudio status=running with no process, so the UI sat on a silent event stream. Persist the worker pid and disable proxy buffering on the SSE response.
Co-authored-by: Cursor <cursoragent@cursor.com>
content=null is rejected even when tool_calls is set, which 400d the U19 follow-up turn after the first graph query.
Co-authored-by: Cursor <cursoragent@cursor.com>
Resume reused the previous pipeline_complete, so the UI closed the stream and showed no live log while U19 was still reviewing.
Co-authored-by: Cursor <cursoragent@cursor.com>
Resume reuses pintables, skips LCSC/DigiKey before review, and emits a waiting heartbeat so a long DeepSeek turn is visible instead of a freeze.
Co-authored-by: Cursor <cursoragent@cursor.com>
Empty assistant content is still treated as unset; replay a placeholder and retry once with thinking disabled.
Co-authored-by: Cursor <cursoragent@cursor.com>
Successful reviews are kept; skipped ICs such as a DeepSeek 400 are run again. Reprocess all re-reviews every IC while still using the library cache.
Co-authored-by: Cursor <cursoragent@cursor.com>
A thinking-only turn was replayed as content=null with no tool_calls, which the API rejects with 400 and skipped IC review.
Co-authored-by: Cursor <cursoragent@cursor.com>
Decode EIA/Yageo/AVX chip codes into library specs, and find datasheets when the file is stored under the orderable catalog name.
Co-authored-by: Cursor <cursoragent@cursor.com>
Review skipped U* when the PDF was stored under a catalog MPN or the part was only in the netlist. Lookup now uses spelling variants and aliases.
Co-authored-by: Cursor <cursoragent@cursor.com>
Series datasheets were stalling extract_pattern on DeepSeek vision; catalog specs are enough for chip R/C/L. Pattern PDF is a 90s fallback only.
Co-authored-by: Cursor <cursoragent@cursor.com>
Store BOM MPN and orderable code on the same PDF blob, derive TI datasheet slugs from package/Q1 codes, and map LCSC/DigiKey R/C/L parameters to specs without a model call.
Co-authored-by: Cursor <cursoragent@cursor.com>
Auto-resolve of passives was 400ing because thinking mode cannot be combined with named tool_choice. Also strip BOM descriptions after an em dash before DigiKey search.
Co-authored-by: Cursor <cursoragent@cursor.com>
micheleandCursor